File size: 14,581 Bytes
ec50e73
792e854
92f5f56
 
 
 
 
 
 
488a53d
ec50e73
 
 
dc6da18
ec50e73
 
 
 
dc6da18
 
ec50e73
792e854
ec50e73
 
b17d0a0
 
ec50e73
dc6da18
a7d545f
 
f40456d
db79ccc
 
f84e07f
 
 
 
 
ec50e73
 
4273b28
ec50e73
 
 
0113778
 
ec50e73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0113778
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ec50e73
 
0113778
ec50e73
 
 
 
 
adf07d5
 
 
ec50e73
 
 
adf07d5
dc6da18
ec50e73
 
adf07d5
ec50e73
4273b28
 
ec50e73
 
adf07d5
4273b28
 
dc6da18
 
 
 
 
 
adf07d5
 
4273b28
 
adf07d5
 
4273b28
 
 
 
dc6da18
 
4273b28
 
 
dc6da18
 
ec50e73
 
adf07d5
4273b28
dc6da18
 
4273b28
 
 
dc6da18
 
ec50e73
dc6da18
4273b28
dc6da18
 
4273b28
 
 
dc6da18
 
4273b28
adf07d5
4273b28
adf07d5
 
dc6da18
 
 
 
 
 
 
 
 
adf07d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ec50e73
adf07d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4273b28
dc6da18
 
 
4273b28
 
adf07d5
 
 
 
 
 
dc6da18
adf07d5
 
dc6da18
adf07d5
 
 
 
 
ec50e73
dc6da18
adf07d5
dc6da18
 
 
 
 
 
 
 
adf07d5
 
 
 
 
 
 
 
 
 
 
 
 
 
84228e7
adf07d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ec50e73
 
91e4b1d
 
2b819e0
91e4b1d
742ceb6
91e4b1d
 
 
9c03b9a
91e4b1d
ec50e73
 
 
adf07d5
ec50e73
 
 
 
 
 
 
adf07d5
 
 
 
ec50e73
adf07d5
 
 
ec50e73
3edb153
 
 
 
 
 
 
 
 
adf07d5
 
 
 
ec50e73
1bfcc16
adf07d5
ec50e73
91e4b1d
1bfcc16
 
adf07d5
 
ec50e73
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
import os
import subprocess

# Install flash attention
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)


import copy
import spaces
import time
import torch

from threading import Thread
from typing import List, Dict, Union
from urllib.parse import urlparse
from PIL import Image

import gradio as gr
from transformers import AutoProcessor, TextIteratorStreamer
from transformers import Idefics2ForConditionalGeneration


DEVICE = torch.device("cuda")
MODELS = {
    "idefics2-8b (sft)": Idefics2ForConditionalGeneration.from_pretrained(
        "HuggingFaceM4/idefics2-8b",
        torch_dtype=torch.bfloat16,
        _attn_implementation="flash_attention_2",
        trust_remote_code=True,
        token=os.environ["HF_AUTH_TOKEN"],
    ).to(DEVICE),
    "idefics2-8b-chatty (chat)": Idefics2ForConditionalGeneration.from_pretrained(
        "HuggingFaceM4/idefics2-8b-chatty",
        torch_dtype=torch.bfloat16,
        _attn_implementation="flash_attention_2",
        trust_remote_code=True,
        token=os.environ["HF_AUTH_TOKEN"],
    ).to(DEVICE),
}
PROCESSOR = AutoProcessor.from_pretrained(
    "HuggingFaceM4/idefics2-8b",
    token=os.environ["HF_AUTH_TOKEN"],
)

# SYSTEM_PROMPT = [ # Deactivating the system propmpt for now, but if I were to reactivate it, I would need to a/ transform turns into dict for applying the chat template, b/ manually overwrite the `default_template` to add the first line (that is not part of any turns), in particular for handling the bos_token.
# #     """The following is a conversation between a highly knowledgeable and intelligent visual AI assistant, called Assistant, and a human user, called User. In the following interactions, User and Assistant will converse in natural language, and Assistant will do its best to answer User’s questions. Assistant has the ability to perceive images and reason about the content of visual inputs. Assistant was built to be respectful, polite and inclusive. It knows a lot, and always tells the truth. When prompted with an image, it does not make up facts.

# The conversation begins:""",
#     """\nUser:""",
#     "https://i1.pickpik.com/photos/515/906/355/kittens-cats-pet-cute-preview.jpg",
#     "Describe this image.<end_of_utterance>",
#     """\nAssistant: Five kittens are standing together in the center of the photograph. From the left to right, there is one orange kitten, two kittens white and black stripes, and two brown kittens with an orange head. They are in bright green grass and it looks like they are moving forward.<end_of_utterance>""",
#     "\nUser:How about this image?",
#     "https://cdn.pixabay.com/photo/2017/09/25/13/12/puppy-2785074_1280.jpg",
#     "Can you describe it too?<end_of_utterance>",
#     """\nAssistant: A dog is lying on the floor, looking at the camera. It is looking directly at you. The dog has a white body and brown patches on its face and ears. Its eyes are dark. Its nose is black, and it has long, floppy ears, white paws, long fur, big eyes.<end_of_utterance>""",
#     "\nUser: What kind of breed is it?<end_of_utterance>",
#     """\nAssistant: It looks like a Cavalier King Charles Spaniel. These breeds are known for their gentle, affectionate nature. Their long ears and big, brown eyes give them a sweet, expressive face. Cavaliers are small dogs weighing between 13 and 20 pounds. They have a silky, medium-length coat that comes in various colors, including black and tan, tricolor, red and white, or black.<end_of_utterance>""",
#     "\nUser: What can you tell me about this breed of dogs?<end_of_utterance>",
#     """\nAssistant: One specific characteristic of the Cavalier King Charles Spaniel is its friendly and affectionate nature. This breed is known for being extremely sociable and forming strong bonds with their owners. They are often described as "velcro dogs" because they love to be close to their human companions, whether it's sitting on their laps, cuddling on the couch, or simply following them around the house.<end_of_utterance>""",
#     "\nUser: ghjkhjabnufs<end_of_utterance>",
#     """\nAssistant: That doesn’t seem to be a word. Could you ask me another way?<end_of_utterance>""",
#     "\nUser: Do you like Cavalier King Charles Spaniel?<end_of_utterance>",
#     """\nAssistant: I do not have personal opinions as I’m just a computer program. However, cavaliers are known for being excellent family pets due to their gentle and patient demeanor, making them great with children and other pets. Their affectionate nature also makes them ideal therapy dogs, as they can provide comfort and emotional support to those in need.<end_of_utterance>""",
#     "\nUser: How many dogs do you see in this image?",
#     "https://i.dailymail.co.uk/i/pix/2011/07/01/article-2010308-0CD22A8300000578-496_634x414.jpg",
#     "<end_of_utterance>",
#     """\nAssistant: There is no dogs in this image. The picture shows a tennis player jumping to volley the ball.<end_of_utterance>""",
# ]

SYSTEM_PROMPT = [
    {
        "role": "user",
        "content": [
             {"type": "image", "image": "https://i1.pickpik.com/photos/515/906/355/kittens-cats-pet-cute-preview.jpg"},
             {"type": "text", "text": "Describe this image."},
        ],
    },
    {
        "role": "assistant",
        "content": [
            {"type": "text", "text": "Five kittens are standing together in the center of the photograph. From the left to right, there is one orange kitten, two kittens white and black stripes, and two brown kittens with an orange head. They are in bright green grass and it looks like they are moving forward."},
        ],
    },
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "How about this image?"},
            {"type": "image", "image": "https://cdn.pixabay.com/photo/2017/09/25/13/12/puppy-2785074_1280.jpg"},
            {"type": "text", "text": "Can you describe it too?"},
        ],
    },
    {
        "role": "assistant",
        "content": [
                {"type": "text", "text": "A dog is lying on the floor, looking at the camera. It is looking directly at you. The dog has a white body and brown patches on its face and ears. Its eyes are dark. Its nose is black, and it has long, floppy ears, white paws, long fur, big eyes."},
        ],
    },
]


API_TOKEN = os.getenv("HF_AUTH_TOKEN")
# IDEFICS_LOGO = "https://huggingface.co/spaces/HuggingFaceM4/idefics_playground/resolve/main/IDEFICS_logo.png"
BOT_AVATAR = "IDEFICS_logo.png"


# Chatbot utils
def turn_is_pure_media(turn):
    return turn[1] is None


def format_user_prompt_with_im_history_and_system_conditioning(
    user_prompt, chat_history
) -> List[Dict[str, Union[List, str]]]:
    """
    Produces the resulting list that needs to go inside the processor.
    It handles the potential image(s), the history and the system conditionning.
    """
    resulting_messages = copy.deepcopy(SYSTEM_PROMPT)
    resulting_images = []

    # Format history
    for turn in chat_history:
        if not resulting_messages or (resulting_messages and resulting_messages[-1]["role"] != "user"):
            resulting_messages.append(
                {
                    "role": "user",
                    "content": [],
                }
            )

        if turn_is_pure_media(turn):
            media = turn[0][0]
            resulting_messages[-1]["content"].append({"type": "image"})
            resulting_images.append(Image.open(media))
        else:
            user_utterance, assistant_utterance = turn
            resulting_messages[-1]["content"].append(
                {"type": "text", "text": user_utterance.strip()}
            )
            resulting_messages.append(
                {
                    "role": "assistant",
                    "content": [
                        {"type": "text", "text": user_utterance.strip()}
                    ]
                }
            )

    # Format current input
    if not user_prompt["files"]:
        resulting_messages.append(
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": user_prompt['text']}
                ],
            }
        )
    else:
        # Choosing to put the image first (i.e. before the text), but this is an arbiratrary choice.
        resulting_messages.append(
            {
                "role": "user",
                "content": [{"type": "image"}] * len(user_prompt['files']) + [
                    {"type": "text", "text": user_prompt['text']}
                ]
            }
        )
        resulting_images.extend([Image.open(im['path']) for im in user_prompt['files']])

    return resulting_messages, resulting_images


def extract_images_from_msg_list(msg_list):
    all_images = []
    for msg in msg_list:
        for c_ in msg["content"]:
            if isinstance(c_, Image.Image):
                all_images.append(c_)
    return all_images


@spaces.GPU(duration=180)
def model_inference(
    user_prompt,
    chat_history,
    model_selector,
    decoding_strategy,
    temperature,
    max_new_tokens,
    repetition_penalty,
    top_p,
):
    if user_prompt["text"].strip() == "" and not user_prompt["files"]:
        gr.Error("Please input a query and optionally image(s).")

    if user_prompt["text"].strip() == "" and user_prompt["files"]:
        gr.Error("Please input a text query along the image(s).")

    for file in user_prompt["files"]:
        if not file["mime_type"].startswith("image/"):
            gr.Error("Idefics2 only supports images. Please input a valid image.")

    streamer = TextIteratorStreamer(
        PROCESSOR.tokenizer,
        skip_prompt=True,
        timeout=5.,
    )

    # Common parameters to all decoding strategies
    # This documentation is useful to read: https://huggingface.co/docs/transformers/main/en/generation_strategies
    generation_args = {
        "max_new_tokens": max_new_tokens,
        "repetition_penalty": repetition_penalty,
        "streamer": streamer,
    }

    assert decoding_strategy in [
        "Greedy",
        "Top P Sampling",
    ]
    if decoding_strategy == "Greedy":
        generation_args["do_sample"] = False
    elif decoding_strategy == "Top P Sampling":
        generation_args["temperature"] = temperature
        generation_args["do_sample"] = True
        generation_args["top_p"] = top_p

    # Creating model inputs
    resulting_text, resulting_images = format_user_prompt_with_im_history_and_system_conditioning(
        user_prompt=user_prompt,
        chat_history=chat_history,
    )
    prompt = PROCESSOR.apply_chat_template(resulting_text, add_generation_prompt=True)
    inputs = PROCESSOR(text=prompt, images=resulting_images if resulting_images else None, return_tensors="pt")
    inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
    generation_args.update(inputs)

    # # The regular non streaming generation mode
    # _ = generation_args.pop("streamer")
    # generated_ids = MODELS[model_selector].generate(**generation_args)
    # generated_text = PROCESSOR.batch_decode(generated_ids[:, generation_args["input_ids"].size(-1): ], skip_special_tokens=True)[0]
    # return generated_text

    # The streaming generation mode
    thread = Thread(
        target=MODELS[model_selector].generate,
        kwargs=generation_args,
    )
    thread.start()

    print("Start generating")
    acc_text = ""
    for text_token in streamer:
        time.sleep(0.04)
        acc_text += text_token
        if acc_text.endswith("<end_of_utterance>"):
            acc_text = acc_text[:-18]
        yield acc_text
    print("Success - generated the following text:", acc_text)
    print("-----")


# Hyper-parameters for generation
max_new_tokens = gr.Slider(
    minimum=8,
    maximum=1024,
    value=512,
    step=1,
    interactive=True,
    label="Maximum number of new tokens to generate",
)
repetition_penalty = gr.Slider(
    minimum=0.01,
    maximum=5.0,
    value=1.1,
    step=0.01,
    interactive=True,
    label="Repetition penalty",
    info="1.0 is equivalent to no penalty",
)
decoding_strategy = gr.Radio(
    [
        "Greedy",
        "Top P Sampling",
    ],
    value="Greedy",
    label="Decoding strategy",
    interactive=True,
    info="Higher values is equivalent to sampling more low-probability tokens.",
)
temperature = gr.Slider(
    minimum=0.0,
    maximum=5.0,
    value=0.4,
    step=0.1,
    interactive=True,
    label="Sampling temperature",
    info="Higher values will produce more diverse outputs.",
)
top_p = gr.Slider(
    minimum=0.01,
    maximum=0.99,
    value=0.8,
    step=0.01,
    interactive=True,
    label="Top P",
    info="Higher values is equivalent to sampling more low-probability tokens.",
)


chatbot = gr.Chatbot(
    label="Idefics2",
    avatar_images=[None, BOT_AVATAR],
    height=450,
)


with gr.Blocks(fill_height=True, css=""".gradio-container .avatar-container {height: 40px width: 40px !important;}""") as demo:
    # model selector should be set to `visbile=False` ultimately
    with gr.Row(elem_id="model_selector_row"):
        model_selector = gr.Dropdown(
            choices=MODELS.keys(),
            value=list(MODELS.keys())[0],
            interactive=True,
            show_label=False,
            container=False,
            label="Model",
            visible=True,
        )

    decoding_strategy.change(
        fn=lambda selection: gr.Slider(
            visible=(
                selection in ["contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k"]
            )
        ),
        inputs=decoding_strategy,
        outputs=temperature,
    )
    decoding_strategy.change(
        fn=lambda selection: gr.Slider(
            visible=(
                selection in ["contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k"]
            )
        ),
        inputs=decoding_strategy,
        outputs=repetition_penalty,
    )
    decoding_strategy.change(
        fn=lambda selection: gr.Slider(visible=(selection in ["Top P Sampling"])),
        inputs=decoding_strategy,
        outputs=top_p,
    )

    gr.ChatInterface(
        fn=model_inference,
        chatbot=chatbot,
        # examples=[{"text": "hello"}, {"text": "hola"}, {"text": "merhaba"}],
        title="Idefics2 Playground",
        multimodal=True,
        additional_inputs=[model_selector, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p],
    )

demo.launch()